智能计算系统实验——基于三层神经网络实现手写数字分类

参考：(12条消息) 智能计算系统实验2_名字不如叫摸鱼的博客-CSDN博客

文件结构

|-- main_exp_2_1. py 主文件
|-- readme.txt
`-- stu_upload 具体的实现文件夹，填空就在这个文件夹
|-- __init__. py
|-- layers_1. py 全连接层、relu 层、softmax 层实现
`-- mnist_mlp_cpu. py 组网

代码

Layers_1 . py

函数解释

• np. random. normal：从正态（高斯）分布中抽取随机样本, loc 表示均值，scale 表示均值，size 表示输出的维度
• np.zeros：返回来一个给定形状和类型的用 0 填充的数组
• np.dot ：矩阵乘法
• axis 参数等于 0 的时候表示压缩行，等于 1 的时候压缩列
• Keepdims 表示输出是否保持原来的维度

代码详解

注意维度要对齐，relu 层和 softmax 层没有需要更新的参数

# coding=utf-8
import numpy as np
import struct
import os
import time

class FullyConnectedLayer(object):
    def __init__(self, num_input, num_output):  # 全连接层初始化
        self.num_input = num_input
        self.num_output = num_output
        print('\tFully connected layer with input %d, output %d.' % (self.num_input, self.num_output))
    def init_param(self, std=0.01):  # 参数初始化
        self.weight = np.random.normal(loc=0.0, scale=std, size=(self.num_input, self.num_output))
        self.bias = np.zeros([1, self.num_output])
        #注意偏置初始化为零
    def forward(self, input):  # 前向传播计算
        start_time = time.time()
        self.input = input
        # TODO：全连接层的前向传播，计算输出结果
        #Y = XW + b (2.3)
        # 注意这里矩阵乘法的顺序是先X后W
        self.output = np.dot(self.input, self.weight) + self.bias
        return self.output
    def backward(self, top_diff):  # 反向传播的计算
        # TODO：全连接层的反向传播，计算参数梯度和本层损失
        #(2.4)
        #注意d_weight的维度应该是（m, n），所以矩阵是X的转置乘top_diff
        self.d_weight = np.dot(self.input.T, top_diff)
        #书上实际上top_diff应该是（1, num_output），但是因为工程上算梯度实际上是批量进行计算，所以实际上维度是(batch_size, num_output)
        self.batch_size = top_diff.shape[0]
        #为了保证维度的正确，(1, batch_size)和(batch_size, num_output)相乘
        self.d_bias = np.dot(np.ones(shape=(1,self.batch_size)),top_diff)
        bottom_diff = np.dot(top_diff, self.weight.T)
        return bottom_diff
    def update_param(self, lr):  # 参数更新
        # TODO：对全连接层参数利用参数进行更新
        self.weight = self.weight - lr * self.d_weight
        self.bias = self.bias - lr * self.d_bias
    def load_param(self, weight, bias):  # 参数加载
        assert self.weight.shape == weight.shape
        assert self.bias.shape == bias.shape
        self.weight = weight
        self.bias = bias
    def save_param(self):  # 参数保存
        return self.weight, self.bias

class ReLULayer(object):
    def __init__(self):
        print('\tReLU layer.')
    def forward(self, input):  # 前向传播的计算
        start_time = time.time()
        self.input = input
        # TODO：ReLU层的前向传播，计算输出结果
        #(2.5)
        output = np.maximum(0, self.input)
        return output
    def backward(self, top_diff):  # 反向传播的计算
        # TODO：ReLU层的反向传播，计算本层损失
        #(2.6)
        bottom_diff = top_diff
        bottom_diff[self.input < 0] = 0
        return bottom_diff

class SoftmaxLossLayer(object):
    def __init__(self):
        print('\tSoftmax loss layer.')
    def forward(self, input):  # 前向传播的计算
        # TODO：softmax 损失层的前向传播，计算输出结果
        #(2.11)
        #input的维度是(batch_size, num_input)
        input_max = np.max(input, axis=1, keepdims=True) #input_max维度是(batch_size, 1)
        input_exp = np.exp(input - input_max)#input_exp维度是(batch_size, num_input)
        #注意np.sum的返回值是一维数组，所以需要加上keepdims参数
        self.prob = input_exp / np.sum(input_exp, axis=1, keepdims=True) 
        return self.prob
    def get_loss(self, label):   # 计算损失
        #self.prob的维度是(batch_size, num_input)
        #(2.12)
        self.batch_size = self.prob.shape[0]
        self.label_onehot = np.zeros_like(self.prob)
        self.label_onehot[np.arange(self.batch_size), label] = 1.0 #batch_size行，label列变成1
        loss = -np.sum(np.log(self.prob) * self.label_onehot) / self.batch_size
        return loss
    def backward(self):  # 反向传播的计算
        # TODO：softmax 损失层的反向传播，计算本层损失
        #(2.13)
        bottom_diff = (self.prob - self.label_onehot) / self.batch_size 
        return bottom_diff

mnist_mlp_cpu. py

函数解释

• struct.unpack_from：按照格式字符串从偏移处开始解包文件，参数分别是格式字符串、解包文件、偏移
  • 格式字符串 https://www.cnblogs.com/flydean/p/14665510.html#%E5%AD%97%E8%8A%82%E9%A1%BA%E5%BA%8F%E5%A4%A7%E5%B0%8F%E5%92%8C%E5%AF%B9%E9%BD%90%E6%96%B9%E5%BC%8F
• np.random.shuffle：打乱数组，没有返回值
• np.argmax：返回元素最大值所对应的索引值

代码详解

# coding=utf-8
import numpy as np
import struct
import os
import time

from layers_1 import FullyConnectedLayer, ReLULayer, SoftmaxLossLayer

MNIST_DIR = "../mnist_data"
TRAIN_DATA = "train-images-idx3-ubyte"
TRAIN_LABEL = "train-labels-idx1-ubyte"
TEST_DATA = "t10k-images-idx3-ubyte"
TEST_LABEL = "t10k-labels-idx1-ubyte"


class MNIST_MLP(object):
    def __init__(self, batch_size=1000, input_size=784, hidden1=32, hidden2=16, out_classes=10, lr=0.01, max_epoch=1, print_iter=100):
        self.batch_size = batch_size
        self.input_size = input_size
        self.hidden1 = hidden1
        self.hidden2 = hidden2
        self.out_classes = out_classes
        self.lr = lr
        self.max_epoch = max_epoch
        self.print_iter = print_iter

    def load_mnist(self, file_dir, is_images = 'True'):
        # Read binary data
        bin_file = open(file_dir, 'rb')
        bin_data = bin_file.read()
        bin_file.close()
        # Analysis file header
        if is_images:
            # Read images
            fmt_header = '>iiii' #>表示大端序，i表示int，四个字节，对应数据集的格式
            magic, num_images, num_rows, num_cols = struct.unpack_from(fmt_header, bin_data, 0)
        else:
            # Read labels
            fmt_header = '>ii'
            magic, num_images = struct.unpack_from(fmt_header, bin_data, 0)
            num_rows, num_cols = 1, 1
        data_size = num_images * num_rows * num_cols
        #B表示1个字节，struct.calcsize返回与格式字符串 format 相对应的结构的大小，也就是跳过了数据集文件的文件头
        mat_data = struct.unpack_from('>' + str(data_size) + 'B', bin_data, struct.calcsize(fmt_header))
        mat_data = np.reshape(mat_data, [num_images, num_rows * num_cols])
        #这里吧mat_data拉成了每行表示一个图片的结构
        print('Load images from %s, number: %d, data shape: %s' % (file_dir, num_images, str(mat_data.shape)))
        return mat_data

    def load_data(self):
        # TODO: 调用函数 load_mnist 读取和预处理 MNIST 中训练数据和测试数据的图像和标记
        print('Loading MNIST data from files...')
        train_images = self.load_mnist(os.path.join(MNIST_DIR, TRAIN_DATA), True)
        train_labels = self.load_mnist(os.path.join(MNIST_DIR, TRAIN_LABEL), False)
        test_images = self.load_mnist(os.path.join(MNIST_DIR, TEST_DATA), True)
        test_labels = self.load_mnist(os.path.join(MNIST_DIR, TEST_LABEL), False)
        self.train_data = np.append(train_images, train_labels, axis=1)
        self.test_data = np.append(test_images, test_labels, axis=1)

    def shuffle_data(self):
        print('Randomly shuffle MNIST data...')
        np.random.shuffle(self.train_data)

    def build_model(self):  # 建立网络结构
        # TODO：建立三层神经网络结构
        print('Building multi-layer perception model...')
        self.fc1 = FullyConnectedLayer(self.input_size, self.hidden1)
        self.relu1 = ReLULayer()
        self.fc2 = FullyConnectedLayer(self.hidden1, self.hidden2)
        self.relu2 = ReLULayer()
        self.fc3 = FullyConnectedLayer(self.hidden2, self.out_classes)
        self.softmax = SoftmaxLossLayer()
        self.update_layer_list = [self.fc1, self.fc2, self.fc3]

    def init_model(self):
        print('Initializing parameters of each layer in MLP...')
        for layer in self.update_layer_list:
            layer.init_param()

    def load_model(self, param_dir):
        print('Loading parameters from file ' + param_dir)
        params = np.load(param_dir).item()
        self.fc1.load_param(params['w1'], params['b1'])
        self.fc2.load_param(params['w2'], params['b2'])
        self.fc3.load_param(params['w3'], params['b3'])

    def save_model(self, param_dir):
        print('Saving parameters to file ' + param_dir)
        params = {}
        params['w1'], params['b1'] = self.fc1.save_param()
        params['w2'], params['b2'] = self.fc2.save_param()
        params['w3'], params['b3'] = self.fc3.save_param()
        np.save(param_dir, params)

    def forward(self, input):  # 神经网络的前向传播
        # TODO：神经网络的前向传播
        h1 = self.fc1.forward(input)
        h1 = self.relu1.forward(h1)
        h2 = self.fc2.forward(h1)
        h2 = self.relu2.forward(h2)
        h3 = self.fc3.forward(h2)
        prob = self.softmax.forward(h3)
        return prob

    def backward(self):  # 神经网络的反向传播
        # TODO：神经网络的反向传播
        dloss = self.softmax.backward()
        dh3 = self.fc3.backward(dloss)
        dh2 = self.relu2.backward(dh3)
        dh2 = self.fc2.backward(dh2)
        dh1 = self.relu1.backward(dh2)
        dh1 = self.fc1.backward(dh1)

    def update(self, lr):
        for layer in self.update_layer_list:
            layer.update_param(lr)

    def train(self):
        max_batch = self.train_data.shape[0] / self.batch_size
        print('Start training...')
        for idx_epoch in range(self.max_epoch):
            self.shuffle_data() # 每个epoch都要进行打乱
            for idx_batch in range(max_batch):
                batch_images = self.train_data[idx_batch*self.batch_size:(idx_batch+1)*self.batch_size, :-1]#切片不包含第二个参数表示的那一行（列）
                batch_labels = self.train_data[idx_batch*self.batch_size:(idx_batch+1)*self.batch_size, -1] #-1表示最后一列
                prob = self.forward(batch_images)
                loss = self.softmax.get_loss(batch_labels)
                self.backward()
                self.update(self.lr)
                if idx_batch % self.print_iter == 0:
                    print('Epoch %d, iter %d, loss: %.6f' % (idx_epoch, idx_batch, loss))

    def evaluate(self):
        pred_results = np.zeros([self.test_data.shape[0]])
        for idx in range(self.test_data.shape[0]/self.batch_size):
            batch_images = self.test_data[idx*self.batch_size:(idx+1)*self.batch_size, :-1]
            start = time.time()
            prob = self.forward(batch_images)# prob维度是(batch_size, 10)
            end = time.time()
            print("inferencing time: %f"%(end-start))
            pred_labels = np.argmax(prob, axis=1) # pred_labels维度是(batch_size, 1)
            pred_results[idx*self.batch_size:(idx+1)*self.batch_size] = pred_labels
        accuracy = np.mean(pred_results == self.test_data[:,-1])
        print('Accuracy in test set: %f' % accuracy)

def build_mnist_mlp(param_dir='weight.npy'):
    h1, h2, e = 320, 160, 200
    mlp = MNIST_MLP(hidden1=h1, hidden2=h2, max_epoch=e)
    mlp.load_data()
    mlp.build_model()
    mlp.init_model()
    mlp.train()
    mlp.save_model('mlp-%d-%d-%depoch.npy' % (h1, h2, e))
    # mlp.load_model('mlp-%d-%d-%depoch.npy' % (h1, h2, e))
    return mlp

if __name__ == '__main__':
    mlp = build_mnist_mlp()
    mlp.evaluate()

main_exp_2_1. py

该文件中的功能和 mnist_mlp_cpu. py 中的 evaluate 函数基本相同，不赘述。